Confirmed
Gemini gains agentic video understanding, cutting token use up to 88% by choosing what to watch
Google announced agentic video understanding for Gemini on September 1, 2026: Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite now choose what to watch, at what speed, and through which modality, frames, audio, or transcript, rather than scanning at a fixed frame rate. Google reports up to 88% lower token use, 66% lower cost, and 7% higher accuracy on tasks like moment retrieval and long video search, live now via the Gemini API and Gemini Enterprise Agent Platform.
In plain termsInstead of watching every second of a video at the same fixed rate, the model now skims: it decides which parts to actually look at closely, the way a person scrubs a video timeline to the parts that matter instead of watching start to finish.
Why it mattersFor anyone optimizing video content for AI retrieval, this is the mechanism Google will lean on to make YouTube and web video searchable by moment rather than by whole upload, once the promised 'Ask YouTube' integration ships.
Our takeThe real test of an 88% token-reduction claim like this is whether it holds on video the model has not seen a benchmark built around, since selective-attention systems tend to look great on curated moment-retrieval tasks and mediocre on messy real footage. Watch whether the 'Ask YouTube' rollout ships with the same efficiency numbers, since a consumer product at YouTube's scale is a very different cost problem than an API preview.
Confirmed
ChatGPT for Healthcare adds a live Epic EHR integration and a nine-source public health data plugin
OpenAI announced on September 1, 2026 that ChatGPT for Healthcare can now connect to Epic electronic health record environments, bringing authorized patient context, appointment notes, lab results, medications, and specialist documentation directly into ChatGPT so clinicians can ask what changed since a patient's last visit. A companion Healthcare Public Data plugin adds structured access to nine official sources, including PubMed, DailyMed, and CMS Coverage. OpenAI says the integration supports the controls and compliance healthcare work requires.
In plain termsAn EHR, electronic health record, is the medical chart system doctors already use for a patient's history, so this integration lets ChatGPT read that chart, with permission, and summarize it instead of a clinician manually paging through notes.
Why it mattersFor clinicians the biggest AI-adoption barrier has been reaching real patient records rather than a general chatbot, and this is one of the first credentialed EHR integrations from a major consumer AI vendor to clear that bar in production.
Our takeA credentialed integration into a live records system is a different risk class than a general-purpose assistant, so watch what OpenAI discloses about audit logging and data retention here, not just the feature list. The nine-source public data plugin is the more exportable part of this launch: expect other regulated verticals to get the same authorized-private-data-plus-vetted-public-data pattern next.
Confirmed
Databricks ships Genie Ontology, ranking business context by authority for AI agents accessed over MCP
Databricks announced Genie Ontology on September 1, 2026, a layer that combines modeled business semantics with context learned from governed tables, queries, dashboards, notebooks, and other assets teams already use. The system ranks that context by authority and relevance, applies permissions, and delivers it to Genie at answer time, and external agents can reach it through MCP, the Model Context Protocol. Databricks frames it as a progressive build: data foundations, enriched metadata, modeled semantics, curated assets, governance, and ongoing evaluation.
In plain termsAn ontology here just means a map of what your business terms actually mean and how they connect, revenue to region to product line, so the AI agent stops guessing at definitions and uses the same ones your analysts already agreed on.
Why it mattersFor any team feeding a data warehouse into an agent, the failure mode has always been an agent reasoning confidently from incomplete or stale metadata, and this is Databricks' answer: authority-ranked context enforced with permissions before it reaches the model.
Our takeI wrote in 'An MCP server is a supply-chain dependency with reach into your data' that the protocol standardizes the wire format between a model and a tool, not the tool's trustworthiness, and that the spec itself tells the common stdio case to skip fine-grained scoping and pull credentials straight from the environment. Genie Ontology's authority-ranking and permission layer sitting in front of MCP access is a more disciplined design than that default, but it only holds if the permission check runs on every read, not just the ones a demo happens to show.
Confirmed
Claude Fable 5.1 reaches private preview inside Snowflake Cortex AI
Snowflake announced on September 1, 2026 that Claude Fable 5.1, described as Anthropic's most capable model to date, is available in private preview on Snowflake Cortex AI, running inside Snowflake's secure perimeter and compatible with Cortex Agents, Cortex AI Functions, Cortex Inference, CoCo, and CoWork. Snowflake highlights improvements in long-running agentic coding, agentic knowledge work, and scientific discovery, plus adjustable effort levels that trade reasoning depth against latency and built-in self-verification loops.
In plain termsCortex AI is Snowflake's own AI layer that runs inside the same walled-off environment as your data, so private preview inside the secure perimeter means the model never has to leave the platform to process your information, the way sending data to an outside API would.
Why it mattersPractitioners running analysis inside Snowflake now get frontier Claude reasoning without moving governed data outside the platform's compliance boundary, which was the main blocker to using a hosted model on regulated datasets.
Our takeA model landing inside a governed data platform matters more for adoption than its benchmark score, since the practical blocker for regulated industries has been getting frontier reasoning without the data ever leaving a compliance boundary. Watch what compute pricing Snowflake attaches to this once it leaves private preview, since Cortex's per-token rates have historically run above calling a model directly.
Confirmed
Google ships AI.PREDICT, bringing TabFM's zero-shot predictions into BigQuery SQL
Google Cloud announced on September 1, 2026 that TabFM, its foundation model for tabular regression and classification, is now callable in BigQuery through a new AI.PREDICT SQL function, in preview. Passing labeled historical data and a new prediction table into one function returns predictions with no model training and no manual feature engineering, since TabFM handles missing values and categorical encoding automatically; a companion AI.EVALUATE function generates standard accuracy metrics in one step.
In plain termsZero-shot means the model makes a prediction without ever being trained specifically on your data first, the way an experienced consultant can size up a new problem from general experience instead of needing months of onboarding.
Why it mattersThis is the actual shipping of the AI.PREDICT function Google only promised as 'coming weeks' back in July, so any team that shelved a TabFM pilot waiting for it now has a real preview to test, not a promise.
Our takeI wrote in 'Google TabFM: the honest read on the benchmark' that as of early July there was no AI.PREDICT function, only Google's promise of 'coming weeks,' and that the underlying benchmark leaderboard was co-created by the same teams behind the models TabFM beat. AI.PREDICT has now actually shipped, which resolves the first concern; the leaderboard concern has not moved, so the right test is still whether it beats a trained model on your own data, not whether it beats a competitor's foundation model on someone else's benchmark.
Deep dive: Google's TabFM brings zero-shot prediction to BigQuery, but the benchmark that sells it is a chart in a blog post
Observed
Bing spotted testing search snippets that drop the site name and keep only the URL
Search Engine Roundtable reported and independently replicated a Bing test, spotted first via a screenshot shared on Search Engine Watch, that removes the site name from a search result snippet and shows only the URL. Microsoft has not confirmed the test, so its rollout scope, population, and duration are unknown. The change echoes a similar Google snippet test from roughly a year earlier.
In plain termsA search results page normally shows a brand name like 'Nike' above the link; this test replaces that name with the plain web address instead, so the URL itself has to carry the trust signal a brand name used to.
Why it mattersIf this spreads, brand recognition in the results page shifts from the site name to the raw URL, which changes how much a clean, readable domain and path structure matters for a searcher's snap trust judgment.
Our takeAn unconfirmed interface test is worth tracking, not acting on, since Bing runs constant results-page experiments that never ship broadly. If it does ship, sites with long or parameter-heavy URLs lose a branding cue they currently get for free, which is one more argument for clean, readable URL paths regardless of what Bing does with this particular test.